MiniMax ARR surges 500% and token consumption jumps 2000% as the agent dividend kicks in

MiniMax's latest disclosures show a commercialization surge: as of August 2026, the company's annual recurring revenue (ARR) has surpassed $800 million, up from just $150 million disclosed in February — roughly a 5.3x expansion in half a year. Token consumption has climbed in tandem, reaching 20 times January's level in July, according to QbitAI.

The growth is now showing up in actual revenue. In the first half of 2026, MiniMax generated $116.6 million in revenue, up 283.1% year over year, already exceeding its full-year 2025 total of $79.04 million; second-quarter revenue grew another 81.8% quarter over quarter.

Even more notable than the scale is the shifting revenue mix. B2B revenue accounted for 63.4% of total revenue in H1 2026 and rose to 80% by August, versus roughly 30% for all of 2025. MiniMax's growth engine has clearly moved from consumer products toward enterprises and developers.

In H1 2026, revenue from MiniMax's open platform and other AI enterprise services reached $73.93 million, up 703.1% year over year, contributing about three-quarters of the total revenue increase. The consumer side did not shrink: AI-native product revenue, covering Talkie and Hailuo AI, came to about $42.64 million, up 100.9%.

Financial efficiency is improving too. Gross profit hit $20.81 million in H1, up 464.8%, lifting gross margin from 12.1% to 17.9%, while sales and distribution expenses fell 17.9% to $26.97 million. R&D spending remained heavy at $296.9 million, up 138.8% year over year.

Why now? CEO Yan Junjie said on the earnings call that enterprise customers and developers on the platform surpassed 2 million by May — ten times the level at the end of last year — while July token consumption hit 20x January's level. Agent-driven inference demand is growing far faster than human user counts and message volumes. MiniMax's models are increasingly entering the real workloads of enterprises and developers, steadily converting model capability into API calls and commercial revenue.

Models and open source are the two accelerators. The M3 model released in June gave developers confidence to plug agent workflows into real business scenarios; the H3 video-generation model open-sourced in July drew more than 24 million downloads within three weeks, spawned over 300 derivative models in the community, and pushed many developers toward paid API usage.

MiniMax's strongest model, M3, costs just $1.2 per million tokens, well below the $2.2 median for comparable models. MiniMax argues that high intelligence and cost efficiency are not a trade-off: lowering the cost per unit of intelligence lets customers run more calls, generating more revenue to reinvest in compute. Whether the commercialization acceleration persists and whether B2B volume can fund the next model cycle are the key things to watch.